LoRA adapters for reading dimension crops from engineering drawings. Fine-tuned on the new-drawings-dataset crop subset (task=="crop") using Unsloth on unsloth/Qwen3.5-4B.

Base model

Base: unsloth/Qwen3.5-4B
Method: LoRA (r=16, α=16), vision + language layers
Chat template: built-in Qwen3.5 template (--chat-template none)

Training

Setting	Value
Train examples	2,598 crop samples
Epochs	2
Batch size	4 × 4 grad accum = 16 effective
Learning rate	2e-4 (cosine)
Max length	2048
Precision	bf16 + 4-bit base load
Trainable params	38.8M / 4.58B (0.85%)
Final eval loss	~0.0032

Dataset: 143 pages, 3,206 crops (80/10/10 train/val/test split). Teacher labels from Gemini 3.5 Flash.

Usage

Load the base model + these LoRA weights with Unsloth or PEFT:

python

from unsloth import FastVisionModel

model, processor = FastVisionModel.from_pretrained(
    "akshayballal/drawing-crop-reader-qwen3.5-4b-lora",
    load_in_4bit=True,
)
model = FastVisionModel.for_inference(model)

messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "<crop-reader prompt>"},
]}]
input_text = processor.apply_chat_template(
    messages, add_generation_prompt=True, enable_thinking=False,
)
inputs = processor(image, input_text, add_special_tokens=False, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)

Files

adapter_model.safetensors — LoRA weights (~148 MB)
adapter_config.json — PEFT config
processor_config.json, tokenizer.json, tokenizer_config.json, chat_template.jinja — processor/tokenizer state from training